Guide

Google's new Gemini agent for work: what it is, and how to write prompts for it

Published 12 October 2026

A sunlit desk with an open paper planner covered in blank coloured sticky notes, a pen, a small plant and a closed laptop

Short answerOn 8 October 2026 Google Cloud announced the Gemini agent, a single agent that handles questions, knowledge work, content and coding from one prompt box, inside Gmail, Docs, Sheets, Slides, Drive, Chat and Calendar. You give it a goal rather than steps. Pricing and rollout dates were not published in the coverage we read.

At its Gemini at Work 2026 event on 8 October 2026, Google Cloud announced the Gemini agent. Google Cloud's announcement, as carried by Zawya, describes it as a "universal agent" that covers everything from knowledge work and answering questions to content creation and coding, from one prompt box.

That page is a press release, and Zawya notes that it was not edited or verified. So treat the claims below as Google's description of its own product.

What Google says it does

One agent, many jobs. It answers questions, does knowledge work, creates images and media, and writes and runs code, all through one agent and one API.
It works inside your apps. Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, plus web, iOS, Android, Windows and Mac, the command line, Microsoft 365 and Slack.
You give it objectives, not steps. It's designed to take a goal and run with it, and it can be scheduled or triggered by events.
It runs in the cloud. Long jobs can carry on after you close your laptop. Google says work can run "for hours or days".
Coworker agents. A team can create an agent with its own email address, calendar, Drive storage and directory entry, by describing the role it needs. Colleagues can add it to a chat or tag it in a document comment to request work, and it can see only what has been shared with it.
It picks the model for each job. Today that means Google's Gemini models and Anthropic's Claude models, with others planned.
Governance features. Each agent has its own identity with limited permissions, every action goes into an audit trail attributed to the agent, and agents run in a sandbox.

What we couldn't find

The coverage doesn't say what it costs, which plans or editions include it, or when it reaches everyone. The only date given is the event itself. If you're deciding whether to rely on it at work, ask your Google Workspace admin or account team, and don't plan around a feature you can't yet switch on.

The skill that matters: write goals, not steps

Whatever tool you use, the habit that this kind of agent rewards is writing a clear goal with limits. These prompts work today in Gemini, ChatGPT or Claude, and they're written so that you can paste the same text into an agent later.

prompt — 5 blanks

Goal: [WHAT SHOULD BE TRUE WHEN YOU'RE DONE, e.g. "a one-page briefing for Thursday's client call"].

Use only: [THE FILES / EMAILS / FOLDERS I'VE SHARED].
Do not: [SEND ANYTHING, EDIT ANYTHING, OR CONTACT ANYONE WITHOUT ASKING ME FIRST].

Deliver: [FORMAT, e.g. a Google Doc, max one page, with a 3-line summary on top].
By: [DEADLINE].

Before you start, tell me your plan in 3-5 bullets and list anything you'd need that you don't have.

The last line is the safety valve. An agent that can act on its own should explain its plan before it acts.

A weekly routine

prompt — 1 blank

Every Friday at 4pm, look at my sent mail and calendar for the week and draft a short status update for [MANAGER]: what shipped, what slipped and why, and next week's top three priorities. Put it in a draft, don't send it. Flag any commitment I made in an email that has no matching calendar item or task.

Our weekly status report prompt shows the format in more detail, and the plan my day prompt is a good small-scale version.

Handing work to a coworker agent

prompt — 1 blank

You are a coworker agent for [TEAM]. Your role: [e.g. "prepare meeting packs"].

When I tag you on a document, do this: read the doc, list open questions, and add them as comments. Don't edit the text itself. When I tag you on a calendar invite, draft an agenda and a list of pre-reads in the event description.

Pair it with our meeting agenda prompt and the delegate a task prompt. Delegating to an agent follows the same rules as delegating to a person: clear outcome, clear limits, clear deadline.

Sensible limits to set

Start read-only. Ask it to draft and summarise before you let it send or edit anything.
Share narrowly. If an agent only sees what you share, share the one folder it needs, not the whole drive.
Keep a human on outgoing messages. For follow-ups, our follow-up prompt writes the draft, and you press send.
Check the audit trail if your admin turns on this feature, and review what the agent actually did.

Where to go from here

For Google's other recent releases, see our Gemini 4 Argon explainer and Gemini Canvas presentation prompts. You'll find more ready-made prompts in the Gemini prompts collection, and the prompt builder can turn a rough goal into a full prompt in a few questions.

Quick checklist

Google Cloud announced the Gemini agent on 8 October 2026 at Gemini at Work.
It's one agent for questions, knowledge work, content and code, inside Workspace apps and other channels.
It takes goals, can run on a schedule, and keeps working in the cloud.
Pricing, plans and a general rollout date were not in the coverage we read.
Write a goal, limits, a format and a deadline, and ask for a plan first.
Start read-only, share narrowly, and keep a person in charge of anything that leaves your inbox.